In the context of air traffic management (ATM), the increasing demand for airspace has exerted substantial pressure on existing systems. This growing demand necessitates the urgent development and implementation of enhanced safety measures and optimized airspace efficiency to accommodate escalating traffic volumes. The complexity of air traffic is determined by the real-time flow of traffic within a given airspace, including the number of flight and their relative positions. Due to the constraints of airspace design, flight must alter their trajectories to fit within these boundaries, resulting in ongoing changes in traffic complexity. As a result, precisely measuring the dynamic complexity of air traffic is a challenging task. This thesis aims to explore the difficulties in flight trajectory prediction and air traffic complexity estimation by leveraging machine learning methodologies, Maximum Entropy principle and information geometry.
Predicting flight trajectories is essential in aviation, as it supports efficient air traffic management and ensures safe and smooth flight operations. Traditional prediction methods often fall short in capturing the complex spatio-temporal dependencies and uncertainties specific to the aviation sector. We propose a novel model named BayesCoordLSTM model. This hybrid model integrates coordinate transformation and Bayesian theorem into ConvLSTM models. By combining the spatial features extracted by the Convolutional Neural Network (CNN) with the temporal dependencies identified by Long Short-Term Memory (LSTM), the model significantly enhances trajectory prediction accuracy. The incorporation of Bayesian theorem offers probabilistic trajectory forecasts with confidence levels, while coordinate transformation improves spatial awareness and predictive performance. We use real flight dataset to implement, our results demonstrate the remarkable superiority of the BayesCoordLSTM model over existing methods. Across various flight scenarios and conditions, this model consistently outperforms baseline models in terms of trajectory prediction accuracy. The utilization of Bayesian optimization not only streamlines hyperparameter tuning but also yields promising results in terms of predictive performance enhancement.
In addition, in airspace complexity estimation, reconstructing angular probability density functions is crucial as it helps model and assess the spatial distribution of air traffic flow. This approach helps assess airspace congestion and predict complex traffic patterns, facilitating more effective air traffic management strategies. In this dissertation, we propose to use the Maximum Entropy principle generating detailed traffic complexity maps from trajectory data, leveraging the statistical properties of Fourier coefficients. The method is validated through numerical simulations, demonstrating its ability to accurately represent angular distributions from simple to complex scenarios.
Using historical flight and airspace traffic data, reconstructing angular probability density functions improves predictions of congestion areas and complexity levels. And then, we illustrate a novel methodology for constructing airspace complexity maps is developed using information geometry. By discretizing the airspace and applying the Generalized von Mises distribution (GvM) alongside Fenchel duality, local traffic complexity indices are calculated based on geodesic distances between distributions. The use of symmetrized Kullback-Leibler divergence ensures computational efficiency, enabling practical application in ATM systems. These complexity maps provide critical insights into traffic patterns and systemic responses, supporting strategic management.
In conclusion, this dissertation proposes efficient approaches to solve some problems in ATM, especially in flight trajectory prediction, and air traffic complexity assessment.